SA-SSL-MOS: Self-supervised Learning MOS Prediction with Spectral Augmentation for Generalized Multi-Rate Speech Assessment

Researchers have developed SA-SSL-MOS, a new self-supervised learning model that significantly improves the assessment of speech quality (MOS prediction) across multiple sampling rates from 16 kHz to 48 kHz. The key innovation is a spectrogram-augmented, parallel-branch architecture and a two-step training scheme that incorporates high-frequency data often discarded by current models. This approach solves a major limitation in existing systems and shows substantially improved generalization, especially when multi-rate training data is limited.